{"id":27078,"date":"2025-02-26T19:29:40","date_gmt":"2025-02-26T11:29:40","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=27078"},"modified":"2025-02-26T19:32:14","modified_gmt":"2025-02-26T11:32:14","slug":"deepgemm","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/deepgemm\/","title":{"rendered":"DeepGEMM\uff1a\u9ad8\u6548\u652f\u6301FP8\u77e9\u9635\u8fd0\u7b97\u7684\u5f00\u6e90\u5e93\uff08DeepSeek \u5f00\u6e90\u5468\u7b2c\u4e09\u5929\uff09"},"content":{"rendered":"<h1><span style=\"font-size: 22px;\">\u7efc\u5408\u4ecb\u7ecd<\/span><\/h1>\n<p>DeepGEMM \u662f\u4e00\u4e2a\u7531 <a href=\"https:\/\/www.kdjingpai.com\/ja\/deepseek-chatshena\/\">DeepSeek<\/a> \u56e2\u961f\u5f00\u53d1\u7684\u5f00\u6e90 FP8 GEMM\uff08\u901a\u7528\u77e9\u9635\u4e58\u6cd5\uff09\u5e93\uff0c\u4e13\u6ce8\u4e8e\u63d0\u4f9b\u9ad8\u6548\u7684\u77e9\u9635\u8fd0\u7b97\u652f\u6301\u3002\u5b83\u7279\u522b\u9488\u5bf9 NVIDIA Hopper \u67b6\u6784\u7684 Tensor Core \u8bbe\u8ba1\uff0c\u652f\u6301\u666e\u901a\u77e9\u9635\u8fd0\u7b97\u548c\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff08MoE\uff09\u7684\u5206\u7ec4 GEMM \u64cd\u4f5c\u3002\u8be5\u5e93\u91c7\u7528 CUDA \u7f16\u5199\uff0c\u901a\u8fc7\u8f7b\u91cf\u7ea7\u7684\u5373\u65f6\u7f16\u8bd1\uff08JIT\uff09\u6280\u672f\u5b9e\u73b0\u8fd0\u884c\u65f6\u5185\u6838\u7f16\u8bd1\uff0c\u65e0\u9700\u5b89\u88c5\u65f6\u9884\u7f16\u8bd1\uff0c\u6781\u5927\u7b80\u5316\u4e86\u90e8\u7f72\u6d41\u7a0b\u3002DeepGEMM \u5728\u4fdd\u6301\u7b80\u6d01\u4ee3\u7801\u7684\u540c\u65f6\uff0c\u6027\u80fd\u8868\u73b0\u51fa\u8272\uff0c\u5728 Hopper GPU \u4e0a\u53ef\u8fbe\u5230\u8d85\u8fc7 1350 TFLOPS \u7684 FP8 \u8ba1\u7b97\u80fd\u529b\u3002\u5b83\u4e0d\u4ec5\u9002\u7528\u4e8e\u673a\u5668\u5b66\u4e60\u6a21\u578b\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u52a0\u901f\uff0c\u8fd8\u56e0\u5176\u5f00\u6e90\u6027\u8d28\u548c\u6613\u8bfb\u6027\uff0c\u6210\u4e3a\u5b66\u4e60 FP8 \u77e9\u9635\u4f18\u5316\u7684\u7edd\u4f73\u8d44\u6e90\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-27081\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/e30857ea3c066a3.jpg\" alt=\"\" width=\"1080\" height=\"1442\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/e30857ea3c066a3.jpg 1080w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/e30857ea3c066a3-768x1025.jpg 768w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<p>&#8211;<strong>\u652f\u6301 FP8 \u77e9\u9635\u8fd0\u7b97<\/strong>\uff1a\u63d0\u4f9b\u9ad8\u6548\u7684 FP8 \u901a\u7528\u77e9\u9635\u4e58\u6cd5\uff08GEMM\uff09\uff0c\u9002\u7528\u4e8e\u9ad8\u6027\u80fd\u8ba1\u7b97\u573a\u666f\u3002<br \/>\n&#8211;<strong>MoE \u6a21\u578b\u4f18\u5316<\/strong>\uff1a\u652f\u6301\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\u7684\u5206\u7ec4 GEMM\uff0c\u4ec5\u5bf9 M \u8f74\u5206\u7ec4\uff0c\u9002\u914d\u4e13\u5bb6\u5171\u4eab\u76f8\u540c\u5f62\u72b6\u7684\u573a\u666f\u3002<br \/>\n&#8211;<strong>\u5373\u65f6\u7f16\u8bd1\uff08JIT\uff09<\/strong>\uff1a\u901a\u8fc7\u8fd0\u884c\u65f6\u7f16\u8bd1\u5185\u6838\uff0c\u65e0\u9700\u9884\u7f16\u8bd1\u5373\u53ef\u9002\u914d\u4e0d\u540c\u786c\u4ef6\u73af\u5883\u3002<br \/>\n&#8211;<strong>\u9ad8\u6027\u80fd\u8ba1\u7b97<\/strong>\uff1a\u5728 NVIDIA Hopper GPU \u4e0a\u5b9e\u73b0\u8d85\u8fc7 1350 TFLOPS \u7684 FP8 \u8ba1\u7b97\u541e\u5410\u91cf\u3002<br \/>\n&#8211;<strong>\u7b80\u6d01\u4ee3\u7801\u8bbe\u8ba1<\/strong>\uff1a\u6838\u5fc3\u4ee3\u7801\u7ea6 300 \u884c\uff0c\u6613\u4e8e\u5b66\u4e60\u548c\u4e8c\u6b21\u5f00\u53d1\u3002<br \/>\n&#8211;<strong>\u517c\u5bb9\u6027\u5f3a<\/strong>\uff1a\u652f\u6301\u666e\u901a GEMM \u548c\u5e26\u63a9\u7801\u7684\u5206\u7ec4 GEMM\uff0c\u9002\u914d\u591a\u79cd\u63a8\u7406\u573a\u666f\u3002<br \/>\n&#8211;<strong>\u5f00\u6e90\u514d\u8d39<\/strong>\uff1a\u57fa\u4e8e MIT \u534f\u8bae\u53d1\u5e03\uff0c\u9002\u7528\u4e8e\u7814\u7a76\u548c\u5546\u4e1a\u7528\u9014\u3002<\/p>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<p>DeepGEMM \u662f\u4e00\u4e2a\u4e13\u4e3a\u5f00\u53d1\u8005\u8bbe\u8ba1\u7684\u5f00\u6e90\u77e9\u9635\u8fd0\u7b97\u5e93\uff0c\u4e3b\u8981\u9762\u5411\u5177\u5907\u4e00\u5b9a CUDA \u7f16\u7a0b\u57fa\u7840\u548c\u673a\u5668\u5b66\u4e60\u80cc\u666f\u7684\u7528\u6237\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u7684\u4f7f\u7528\u6307\u5357\uff0c\u5e2e\u52a9\u4f60\u5feb\u901f\u4e0a\u624b\u5e76\u5c06\u5176\u96c6\u6210\u5230\u9879\u76ee\u4e2d\u3002<\/p>\n<h3>\u5b89\u88c5\u6d41\u7a0b<\/h3>\n<p>DeepGEMM \u4e0d\u9700\u8981\u590d\u6742\u7684\u9884\u7f16\u8bd1\u8fc7\u7a0b\uff0c\u53ea\u9700\u51e0\u6b65\u5373\u53ef\u5b8c\u6210\u5b89\u88c5\u548c\u8fd0\u884c\u73af\u5883\u914d\u7f6e\uff1a<br \/>\n1.<strong>\u73af\u5883\u51c6\u5907<\/strong>\uff1a<\/p>\n<ul>\n<li>\u7cfb\u7edf\u8981\u6c42\uff1a\u652f\u6301 NVIDIA Hopper \u67b6\u6784\u7684 GPU\uff08\u5982 H100\uff09\u3002<\/li>\n<li>\u8f6f\u4ef6\u4f9d\u8d56\uff1a\u5b89\u88c5 CUDA Toolkit\uff08\u5efa\u8bae\u7248\u672c 11.8 \u6216\u66f4\u9ad8\uff09\u548c Python\uff083.8+\uff09\u3002<\/li>\n<li>\u786c\u4ef6\u652f\u6301\uff1a\u786e\u4fdd\u4f60\u7684\u8bbe\u5907\u914d\u5907\u81f3\u5c11 40GB \u663e\u5b58\u7684 NVIDIA GPU\u3002<br \/>\n2.<strong>\u514b\u9686\u4ed3\u5e93<\/strong>\uff1a<br \/>\n\u5728\u7ec8\u7aef\u8fd0\u884c\u4ee5\u4e0b\u547d\u4ee4\uff0c\u5c06 DeepGEMM \u4ed3\u5e93\u4e0b\u8f7d\u5230\u672c\u5730\uff1a<\/li>\n<\/ul>\n<pre><code>git clone https:\/\/github.com\/deepseek-ai\/DeepGEMM.git**\r\ncd DeepGEMM**\r\n<\/code><\/pre>\n<ol start=\"3\">\n<li>\u5b89\u88c5\u4f9d\u8d56\uff1a<br \/>\n\u4f7f\u7528 Python \u7684\u5305\u7ba1\u7406\u5de5\u5177\u5b89\u88c5\u6240\u9700\u4f9d\u8d56\uff1a<\/li>\n<\/ol>\n<pre><code>pip install torch numpy\r\n<\/code><\/pre>\n<p>DeepGEMM \u672c\u8eab\u65e0\u9700\u989d\u5916\u7f16\u8bd1\uff0c\u56e0\u4e3a\u5b83\u4f9d\u8d56\u5373\u65f6\u7f16\u8bd1\u6280\u672f\uff0c\u6240\u6709\u5185\u6838\u4f1a\u5728\u8fd0\u884c\u65f6\u81ea\u52a8\u751f\u6210\u3002<br \/>\n4. \u9a8c\u8bc1\u5b89\u88c5\uff1a<br \/>\n\u8fd0\u884c\u63d0\u4f9b\u7684\u6d4b\u8bd5\u811a\u672c\uff0c\u786e\u4fdd\u73af\u5883\u914d\u7f6e\u6b63\u786e\uff1a<\/p>\n<pre><code>python test\/deep_gemm_test.py\r\n<\/code><\/pre>\n<p>\u5982\u679c\u8f93\u51fa\u663e\u793a\u6b63\u5e38\u7684\u77e9\u9635\u8fd0\u7b97\u7ed3\u679c\uff0c\u8bf4\u660e\u5b89\u88c5\u6210\u529f\u3002<\/p>\n<h3>\u4e3b\u8981\u529f\u80fd\u64cd\u4f5c<\/h3>\n<h4>1. \u6267\u884c\u57fa\u672c\u7684 FP8 GEMM \u8fd0\u7b97<\/h4>\n<p>DeepGEMM \u63d0\u4f9b\u4e86\u4e00\u4e2a\u7b80\u5355\u6613\u7528\u7684\u63a5\u53e3\uff0c\u7528\u4e8e\u6267\u884c\u975e\u5206\u7ec4\u7684 FP8 \u77e9\u9635\u4e58\u6cd5\uff1a<\/p>\n<ul>\n<li>\u64cd\u4f5c\u6b65\u9aa4\uff1a\n<ul>\n<li>\u5bfc\u5165\u5e93\u548c\u51fd\u6570\uff1a<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<pre><code>import torch\r\nfrom deep_gemm import gemm_fp8_fp8_bf16_nt\r\n<\/code><\/pre>\n<ul>\n<li>\u51c6\u5907\u8f93\u5165\u6570\u636e\uff08\u77e9\u9635 A \u548c B\uff0c\u5fc5\u987b\u662f FP8 \u683c\u5f0f\uff09\uff1a<\/li>\n<\/ul>\n<pre><code>A = torch.randn(1024, 512, dtype=torch.float8_e4m3fn).cuda()\r\nB = torch.randn(512, 1024, dtype=torch.float8_e4m3fn).cuda()\r\n<\/code><\/pre>\n<ul>\n<li>\u8c03\u7528\u51fd\u6570\u8fdb\u884c\u77e9\u9635\u4e58\u6cd5\uff1a<\/li>\n<\/ul>\n<pre><code>C = gemm_fp8_fp8_bf16_nt(A, B)\r\nprint(C)\r\n<\/code><\/pre>\n<ul>\n<li>\u6ce8\u610f\u4e8b\u9879\uff1a\n<ul>\n<li>\u8f93\u5165\u77e9\u9635\u9700\u4f4d\u4e8e GPU \u4e0a\uff0c\u4e14\u683c\u5f0f\u9700\u4e3a FP8\uff08E4M3 \u6216 E5M2\uff09\u3002<\/li>\n<li>\u8f93\u51fa\u7ed3\u679c\u4e3a BF16 \u683c\u5f0f\uff0c\u9002\u5408\u540e\u7eed\u8ba1\u7b97\u6216\u5b58\u50a8\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>2. \u652f\u6301 MoE \u6a21\u578b\u7684\u5206\u7ec4 GEMM<\/h4>\n<p>\u5bf9\u4e8e\u9700\u8981\u5904\u7406 MoE \u6a21\u578b\u7684\u7528\u6237\uff0cDeepGEMM \u63d0\u4f9b\u4e86\u5206\u7ec4 GEMM \u652f\u6301\uff1a<\/p>\n<ul>\n<li>\u64cd\u4f5c\u6b65\u9aa4\uff1a\n<ul>\n<li>\u5bfc\u5165\u5206\u7ec4 GEMM \u51fd\u6570\uff1a<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<pre><code>from deep_gemm import m_grouped_gemm_fp8_fp8_bf16_nt_contiguous\r\n<\/code><\/pre>\n<ul>\n<li>\u51c6\u5907\u8fde\u7eed\u5e03\u5c40\u7684\u8f93\u5165\u6570\u636e\uff1a<\/li>\n<\/ul>\n<pre><code>A = torch.randn(4096, 512, dtype=torch.float8_e4m3fn).cuda()  # \u591a\u4e2a\u4e13\u5bb6\u7684\u8f93\u5165\u62fc\u63a5\r\nB = torch.randn(512, 1024, dtype=torch.float8_e4m3fn).cuda()\r\ngroup_sizes = [1024, 1024, 1024, 1024]  # \u6bcf\u4e2a\u4e13\u5bb6\u7684 <a href=\"https:\/\/www.kdjingpai.com\/ja\/tokenization\/\">token<\/a> \u6570\r\n<\/code><\/pre>\n<ul>\n<li>\u6267\u884c\u5206\u7ec4 GEMM\uff1a<\/li>\n<\/ul>\n<pre><code>C = m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(A, B, group_sizes)\r\nprint(C)\r\n<\/code><\/pre>\n<ul>\n<li>\u6ce8\u610f\u4e8b\u9879\uff1a\n<ul>\n<li>\u8f93\u5165\u77e9\u9635 A \u7684 M \u8f74\u9700\u6309\u4e13\u5bb6\u5206\u7ec4\u62fc\u63a5\uff0c\u4e14\u6bcf\u4e2a\u5206\u7ec4\u7684\u5927\u5c0f\u9700\u5bf9\u9f50 GEMM M \u5757\u5927\u5c0f\uff08\u53ef\u7528\u00a0<code>get_m_alignment_for_contiguous_layout()<\/code>\u00a0\u83b7\u53d6\uff09\u3002<\/li>\n<li>B \u77e9\u9635\u7684 N \u548c K \u8f74\u9700\u4fdd\u6301\u56fa\u5b9a\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>3. \u63a8\u7406\u9636\u6bb5\u7684\u63a9\u7801\u5206\u7ec4 GEMM<\/h4>\n<p>\u5728\u63a8\u7406\u89e3\u7801\u9636\u6bb5\uff0cDeepGEMM \u652f\u6301\u4f7f\u7528\u63a9\u7801\u7684\u5206\u7ec4 GEMM\uff0c\u9002\u7528\u4e8e\u52a8\u6001 token \u5206\u914d\uff1a<\/p>\n<ul>\n<li>\u64cd\u4f5c\u6b65\u9aa4\uff1a\n<ul>\n<li>\u5bfc\u5165\u63a9\u7801\u5206\u7ec4\u51fd\u6570\uff1a<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<pre><code>from deep_gemm import m_grouped_gemm_fp8_fp8_bf16_nt_masked\r\n<\/code><\/pre>\n<ul>\n<li>\u51c6\u5907\u8f93\u5165\u6570\u636e\u548c\u63a9\u7801\uff1a<\/li>\n<\/ul>\n<pre><code>A = torch.randn(4096, 512, dtype=torch.float8_e4m3fn).cuda()\r\nB = torch.randn(512, 1024, dtype=torch.float8_e4m3fn).cuda()\r\nmask = torch.ones(4096, dtype=torch.bool).cuda()  # \u63a9\u7801\u6307\u793a\u6709\u6548 token\r\n<\/code><\/pre>\n<ul>\n<li>\u6267\u884c\u63a9\u7801\u5206\u7ec4 GEMM\uff1a<\/li>\n<\/ul>\n<pre><code>C = m_grouped_gemm_fp8_fp8_bf16_nt_masked(A, B, mask)\r\nprint(C)\r\n<\/code><\/pre>\n<ul>\n<li>\u6ce8\u610f\u4e8b\u9879\uff1a\n<ul>\n<li>\u63a9\u7801\u7528\u4e8e\u6307\u5b9a\u54ea\u4e9b token \u9700\u8981\u8ba1\u7b97\uff0c\u9002\u5408 CUDA \u56fe\u542f\u7528\u65f6\u7684\u52a8\u6001\u63a8\u7406\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h3>\u7279\u8272\u529f\u80fd\u64cd\u4f5c\u6d41\u7a0b<\/h3>\n<h4>\u9ad8\u6027\u80fd\u4f18\u5316\u4e0e\u8c03\u8bd5<\/h4>\n<p>DeepGEMM \u7684\u6838\u5fc3\u4f18\u52bf\u5728\u4e8e\u5176\u9ad8\u6548\u6027\u548c\u7b80\u6d01\u6027\uff0c\u5f00\u53d1\u8005\u53ef\u4ee5\u901a\u8fc7\u4ee5\u4e0b\u6b65\u9aa4\u8fdb\u4e00\u6b65\u4f18\u5316\u548c\u8c03\u8bd5\uff1a<\/p>\n<ul>\n<li>\u67e5\u770b\u6027\u80fd\u6570\u636e\uff1a<br \/>\n\u5728\u8fd0\u884c\u6d4b\u8bd5\u811a\u672c\u65f6\uff0c\u6dfb\u52a0\u65e5\u5fd7\u8f93\u51fa\u4ee5\u76d1\u63a7 TFLOPS\uff1a<\/p>\n<pre><code>import logging\r\nlogging.basicConfig(level=logging.INFO)\r\nC = gemm_fp8_fp8_bf16_nt(A, B)\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n<p><em>\u8c03\u6574\u53c2\u6570<\/em>*\uff1a<br \/>\n\u6839\u636e\u5177\u4f53\u786c\u4ef6\u8c03\u6574\u5757\u5927\u5c0f\uff08TMA \u53c2\u6570\uff09\u4ee5\u4f18\u5316\u6570\u636e\u79fb\u52a8\u548c\u8ba1\u7b97\u91cd\u53e0\uff0c\u53c2\u8003\u6587\u6863\u4e2d\u7684 test\/ \u6587\u4ef6\u5939\u4e2d\u7684\u793a\u4f8b\u3002<br \/>\n<em>\u5b66\u4e60\u4e0e\u6269\u5c55<\/em>*\uff1a<br \/>\n\u6838\u5fc3\u4ee3\u7801\u4f4d\u4e8e deep_gemm\/gemm_kernel.cu\uff0c\u7ea6 300 \u884c\uff0c\u5f00\u53d1\u8005\u53ef\u76f4\u63a5\u9605\u8bfb\u5e76\u4fee\u6539\u4ee5\u9002\u914d\u81ea\u5b9a\u4e49\u9700\u6c42\u3002<br \/>\n\u4f7f\u7528\u5efa\u8bae<br \/>\n<em>\u786c\u4ef6\u8981\u6c42<\/em>*\uff1a\u76ee\u524d\u4ec5\u652f\u6301 NVIDIA Hopper \u67b6\u6784 GPU\uff0c\u5176\u4ed6\u67b6\u6784\u6682\u672a\u9002\u914d\u3002<br \/>\n<em>\u6587\u6863\u53c2\u8003<\/em>*\uff1a\u8be6\u7ec6\u51fd\u6570\u8bf4\u660e\u548c\u793a\u4f8b\u4ee3\u7801\u53ef\u5728 GitHub \u4ed3\u5e93\u7684 README.md \u548c test\/ \u6587\u4ef6\u5939\u4e2d\u627e\u5230\u3002<br \/>\n<em>\u793e\u533a\u652f\u6301<\/em>*\uff1a\u5982\u9047\u95ee\u9898\uff0c\u53ef\u5728 GitHub Issues \u9875\u9762\u63d0\u4ea4\u53cd\u9988\uff0cDeepSeek \u56e2\u961f\u4f1a\u79ef\u6781\u54cd\u5e94\u3002<br \/>\n\u901a\u8fc7\u4ee5\u4e0a\u6b65\u9aa4\uff0c\u4f60\u53ef\u4ee5\u8f7b\u677e\u5c06 DeepGEMM \u96c6\u6210\u5230\u673a\u5668\u5b66\u4e60\u9879\u76ee\u4e2d\uff0c\u4eab\u53d7\u5176\u9ad8\u6548\u7684 FP8 \u77e9\u9635\u8fd0\u7b97\u80fd\u529b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7efc\u5408\u4ecb\u7ecd DeepGEMM \u662f\u4e00\u4e2a\u7531 DeepSeek \u56e2\u961f\u5f00\u53d1\u7684\u5f00\u6e90 FP8 GEMM\uff08\u901a\u7528\u77e9\u9635\u4e58\u6cd5\uff09\u5e93\uff0c\u4e13\u6ce8\u4e8e\u63d0\u4f9b\u9ad8\u6548\u7684\u77e9\u9635\u8fd0\u7b97\u652f\u6301\u3002\u5b83\u7279\u522b\u9488\u5bf9 NVIDIA Hopper \u67b6\u6784\u7684 Tensor Core \u8bbe\u8ba1\uff0c\u652f\u6301\u666e\u901a\u77e9\u9635\u8fd0\u7b97\u548c\u6df7\u5408\u4e13&#8230;<\/p>\n","protected":false},"author":1,"featured_media":61928,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[230],"class_list":["post-27078","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","tag-aikaiyuanxiangmu"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/27078","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/comments?post=27078"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/27078\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media\/61928"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media?parent=27078"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/categories?post=27078"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/tags?post=27078"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}